Score-Based Symmetry Diffusion Models

• The paper introduces a novel diffusion framework that enforces physical symmetries via group-equivariant networks and force-regularized score matching to improve sampling near criticality.
• The methodology leverages reverse-time stochastic differential equations and probability-flow ODEs to accurately model noised lattice distributions, yielding high effective sample sizes and reduced autocorrelations.
• Practical insights include enhanced sample quality in φ⁴ and U(1) lattice setups, demonstrating that symmetry preservation can serve as a robust inductive bias for scalable lattice simulations.

Searching arXiv for the cited paper and closely related work on diffusion models, score matching, and lattice field theory. Score-based symmetry-preserving diffusion models are generative models that replace or augment conventional Markov Chain Monte Carlo procedures by learning the score field of a noised lattice distribution while enforcing exact physical symmetries in the score network by construction. In lattice quantum field theory, this approach has been developed for two-dimensional $\phi^4$ and ${\rm U}(1)$ theories in the setting of forward and reverse stochastic differential equations, group-equivariant score parameterizations, and a force-regularized score-matching objective (Vega et al., 30 Oct 2025). The resulting framework targets a central numerical obstacle near criticality—critical slowing down—by combining continuous diffusion-based sampling with exact equivariance under global $\mathbb{Z}_2$ reflections, local ${\rm U}(1)$ gauge rotations, and periodic lattice translations $\mathbb{T}$ (Vega et al., 30 Oct 2025). A broader methodological backdrop is provided by score-based generative modeling and diffusion probabilistic models, which established the reverse-SDE and probability-flow formulations used here (Song et al., 2020, Ho et al., 2020).

1. Stochastic formulation and score-based sampling

The diffusion construction begins from a forward stochastic process on fields $\phi_t$,

$$d\phi_t \;=\; f(\phi_t,t)\,dt \;+\;\sigma(t)\,dW_t\,,$$

with two commonly used parameterizations: the variance-preserving (VP) form

$$d\phi_t = -\tfrac12\beta(t)\,\phi_t\,dt \;+\;\sqrt{\beta(t)}\,dW_t$$

and the variance-expanding (VE) form

$d\phi_t = \sigma(t)\,dW_t\,.$

The corresponding density evolution is governed by the Fokker–Planck equation

$$\partial_t p_t(\phi) =\nabla\!\cdot\!\Bigl[-f(\phi,t)\,p_t(\phi) \;+\;\tfrac12\,\sigma^2(t)\,\nabla p_t(\phi)\Bigr]\,,$$

and the learned object is the score

${\rm U}(1)$0

These ingredients are standard in score-based diffusion modeling, and their reverse-time interpretation underlies the lattice-field applications considered here (Vega et al., 30 Oct 2025, Song et al., 2020).

Sampling proceeds through the reverse-time SDE

${\rm U}(1)$1

where one may choose, for example, ${\rm U}(1)$2. In the special case ${\rm U}(1)$3, the reverse dynamics reduce to the deterministic probability-flow ODE,

${\rm U}(1)$4

The paper also presents a predictor–corrector decomposition,

${\rm U}(1)$5

which makes explicit the relation between drift-based transport and stochastic refinement (Vega et al., 30 Oct 2025).

For numerical implementation, the reverse SDE is discretized with Euler–Maruyama,

${\rm U}(1)$6

while the probability-flow ODE may be integrated with Euler or higher-order solvers such as Runge–Kutta when exact likelihoods are required (Vega et al., 30 Oct 2025). In the lattice setting, this stochastic formulation is not merely a reformulation of generative modeling; it defines a sampling mechanism whose fidelity depends directly on whether the learned score respects the symmetries of the underlying action.

2. Equivariance and exact symmetry preservation

A central structural result is Theorem 3.1 of the paper: if the action, or equivalently ${\rm U}(1)$7, is invariant under a group ${\rm U}(1)$8 acting by ${\rm U}(1)$9, then the score transforms contravariantly,

$\mathbb{Z}_2$0

This theorem motivates building exact equivariance into the score network rather than attempting to recover it statistically from training data (Vega et al., 30 Oct 2025).

For $\mathbb{Z}_2$1 symmetry, relevant to $\mathbb{Z}_2$2 theory, the global flip $\mathbb{Z}_2$3 implies

$\mathbb{Z}_2$4

In the simplest $\mathbb{Z}_2$5D setting, this can be realized with an MLP using an odd activation such as $\mathbb{Z}_2$6, yielding an exactly antisymmetric network (Vega et al., 30 Oct 2025). In the two-dimensional $\mathbb{Z}_2$7 architecture, the same symmetry is enforced by antisymmetrizing the output,

$\mathbb{Z}_2$8

For lattice translations $\mathbb{Z}_2$9 on an ${\rm U}(1)$0 torus, equivariance is implemented through circular padding in every convolution, so that translating the input field by one site translates the output score by one site as well (Vega et al., 30 Oct 2025). This is a direct architectural encoding of periodic boundary conditions.

For ${\rm U}(1)$1 gauge symmetry, two exact constructions are described. One works in the angular representation for link variables ${\rm U}(1)$2, where the gauge action ${\rm U}(1)$3 is affine with unit Jacobian, implying that the score ${\rm U}(1)$4 is exactly gauge-invariant. The other constructs the network on gauge-invariant plaquettes ${\rm U}(1)$5 and then distributes the scalar output back to link directions (Vega et al., 30 Oct 2025). In both cases, the symmetry is imposed by design rather than by data augmentation.

This symmetry-preserving strategy is consistent with the broader literature on equivariant generative modeling, but its role in lattice field theory is especially direct: the score approximates a generalized force field, so violating symmetry at the network level would misrepresent the geometry of the target distribution. A plausible implication is that exact equivariance improves not only sample quality but also the stability of reverse integration, because the drift remains confined to symmetry-compatible directions.

3. Network architectures and force-regularized training

For the two-dimensional ${\rm U}(1)$6 theory, the base score model is a U-Net with down- and up-sampling via strided and transposed convolutions, group-norm, SiLU activations, skip connections, and Gaussian Fourier features ${\rm U}(1)$7 to embed the continuous diffusion time ${\rm U}(1)$8 (Vega et al., 30 Oct 2025). Translation equivariance is enforced through circular padding, and ${\rm U}(1)$9 symmetry is enforced by output antisymmetrization. For the $\mathbb{T}$0 gauge theory, a gauge-equivariant U-Net is built on plaquette angles and outputs two channels for the link-direction score (Vega et al., 30 Oct 2025).

Training is based on denoising score matching. The standard objective is

$\mathbb{T}$1

In practice, one samples

$\mathbb{T}$2

and uses the equivalent simple objective, up to constants,

$\mathbb{T}$3

The paper notes that near $\mathbb{T}$4 the weight $\mathbb{T}$5, so the network learns the small-noise limit by extrapolation (Vega et al., 30 Oct 2025).

Lattice quantum field theory provides an additional structure absent from generic diffusion applications: the exact force

$\mathbb{T}$6

This is used to define the augmented objective

$\mathbb{T}$7

The paper describes this as “force-improved” score matching and states that it both anchors the score at $\mathbb{T}$8 and acts as a strong regularizer, improving sample quality (Vega et al., 30 Oct 2025).

The significance of this modification is specific to field-theoretic sampling. In ordinary score-based generative modeling, the score at vanishing noise is inferred indirectly; here it can be constrained by known variational structure. This suggests a close relation between score learning and force learning in lattice systems, with the diffusion model approximating a hierarchy of coarse-to-fine effective forces across noise scales.

4. Scalar-theory results: from 0D $\mathbb{T}$9 to the two-dimensional broken phase

The paper studies a $\phi_t$0D $\phi_t$1 toy model with action

$\phi_t$2

which has $\phi_t$3 symmetry. The trained model is a 2-layer MLP with OddSigmoid. In this setting, the learned score matches the analytic force in both symmetric and broken phases, and sample histograms from the reverse SDE agree with HMC data (Vega et al., 30 Oct 2025). Effective sample size, computed from importance weights as

$\phi_t$4

is reported as $\phi_t$5 in the symmetric phase and $\phi_t$6 in the broken phase (Vega et al., 30 Oct 2025).

The main two-dimensional scalar experiment uses the $\phi_t$7 broken-phase lattice with action

$\phi_t$8

The observables highlighted are $\phi_t$9, $$d\phi_t \;=\; f(\phi_t,t)\,dt \;+\;\sigma(t)\,dW_t\,,$$0, and the Binder cumulant $$d\phi_t \;=\; f(\phi_t,t)\,dt \;+\;\sigma(t)\,dW_t\,,$$1 (Vega et al., 30 Oct 2025). The comparison of “DM (raw),” “DM (reweighted),” “HMC(test),” and “HMC(train)” shows excellent agreement, with all values within statistical errors (Vega et al., 30 Oct 2025).

A central quantitative comparison concerns autocorrelations of the magnetization. The integrated autocorrelation time

$$d\phi_t \;=\; f(\phi_t,t)\,dt \;+\;\sigma(t)\,dW_t\,,$$2

is reported as $$d\phi_t \;=\; f(\phi_t,t)\,dt \;+\;\sigma(t)\,dW_t\,,$$3 for HMC and $$d\phi_t \;=\; f(\phi_t,t)\,dt \;+\;\sigma(t)\,dW_t\,,$$4 for diffusion+Metropolis–Hastings on this lattice (Vega et al., 30 Oct 2025). The paper interprets this as a substantial reduction in autocorrelation under the hybrid diffusion proposal mechanism.

A more detailed ablation isolates the effect of symmetry enforcement:

Symmetry	ESS	$$d\phi_t \;=\; f(\phi_t,t)\,dt \;+\;\sigma(t)\,dW_t\,,$$5
none	0.11(1)	0.9650(8)
$$d\phi_t \;=\; f(\phi_t,t)\,dt \;+\;\sigma(t)\,dW_t\,,$$6	0.24(3)	0.9798(9)
T (circular)	0.35(4)	0.9867(14)
$$d\phi_t \;=\; f(\phi_t,t)\,dt \;+\;\sigma(t)\,dW_t\,,$$7	0.54(3)	0.9921(6)

The same table reports accept rates of $$d\phi_t \;=\; f(\phi_t,t)\,dt \;+\;\sigma(t)\,dW_t\,,$$8, $$d\phi_t \;=\; f(\phi_t,t)\,dt \;+\;\sigma(t)\,dW_t\,,$$9, $$d\phi_t = -\tfrac12\beta(t)\,\phi_t\,dt \;+\;\sqrt{\beta(t)}\,dW_t$$0, and $$d\phi_t = -\tfrac12\beta(t)\,\phi_t\,dt \;+\;\sqrt{\beta(t)}\,dW_t$$1 for the four symmetry settings, respectively (Vega et al., 30 Oct 2025). The factual conclusion drawn in the paper is that symmetry-aware models outperform generic score networks in sample quality, expressivity, and effective sample size.

The force-regularization and integrator study further reports that force regularization raises ESS, for example from $$d\phi_t = -\tfrac12\beta(t)\,\phi_t\,dt \;+\;\sqrt{\beta(t)}\,dW_t$$2 without symmetries and from $$d\phi_t = -\tfrac12\beta(t)\,\phi_t\,dt \;+\;\sqrt{\beta(t)}\,dW_t$$3 with $$d\phi_t = -\tfrac12\beta(t)\,\phi_t\,dt \;+\;\sqrt{\beta(t)}\,dW_t$$4; that $$d\phi_t = -\tfrac12\beta(t)\,\phi_t\,dt \;+\;\sqrt{\beta(t)}\,dW_t$$5 improves from approximately $$d\phi_t = -\tfrac12\beta(t)\,\phi_t\,dt \;+\;\sqrt{\beta(t)}\,dW_t$$6 without symmetries to approximately $$d\phi_t = -\tfrac12\beta(t)\,\phi_t\,dt \;+\;\sqrt{\beta(t)}\,dW_t$$7 with full symmetry plus force regularization; and that acceptance rates rise from approximately $$d\phi_t = -\tfrac12\beta(t)\,\phi_t\,dt \;+\;\sqrt{\beta(t)}\,dW_t$$8 for the fully symmetric model with force regularization (Vega et al., 30 Oct 2025). Higher-order solvers such as RK4 only marginally improve metrics relative to Euler, but at a factor-of-four cost in number of function evaluations (Vega et al., 30 Oct 2025).

5. Gauge-theory construction and results for two-dimensional $$d\phi_t = -\tfrac12\beta(t)\,\phi_t\,dt \;+\;\sqrt{\beta(t)}\,dW_t$$9

For two-dimensional $d\phi_t = \sigma(t)\,dW_t\,.$0 lattice gauge theory on $d\phi_t = \sigma(t)\,dW_t\,.$1 lattices, the action is

$d\phi_t = \sigma(t)\,dW_t\,.$2

where $d\phi_t = \sigma(t)\,dW_t\,.$3 denotes the plaquette angle (Vega et al., 30 Oct 2025). The exact force is given by

$d\phi_t = \sigma(t)\,dW_t\,.$4

This exact expression supplies the same kind of structural information used in the scalar case, but now in a gauge-equivariant setting (Vega et al., 30 Oct 2025).

The score model is a gauge-equivariant U-Net defined on plaquette angles and returning two channels for the link-direction score (Vega et al., 30 Oct 2025). Reverse sampling dynamics are assessed through the average plaquette and topological charge $d\phi_t = \sigma(t)\,dW_t\,.$5, both of which evolve smoothly to HMC values as $d\phi_t = \sigma(t)\,dW_t\,.$6 (Vega et al., 30 Oct 2025). Wilson-loop measurements for rectangular loops up to size $d\phi_t = \sigma(t)\,dW_t\,.$7 match both HMC and exact values (Vega et al., 30 Oct 2025).

The paper also reports the plaquette expectation value $d\phi_t = \sigma(t)\,dW_t\,.$8 and topological susceptibility $d\phi_t = \sigma(t)\,dW_t\,.$9 at $$\partial_t p_t(\phi) =\nabla\!\cdot\!\Bigl[-f(\phi,t)\,p_t(\phi) \;+\;\tfrac12\,\sigma^2(t)\,\nabla p_t(\phi)\Bigr]\,,$$0. For the plaquette, HMC values are approximately $$\partial_t p_t(\phi) =\nabla\!\cdot\!\Bigl[-f(\phi,t)\,p_t(\phi) \;+\;\tfrac12\,\sigma^2(t)\,\nabla p_t(\phi)\Bigr]\,,$$1, $$\partial_t p_t(\phi) =\nabla\!\cdot\!\Bigl[-f(\phi,t)\,p_t(\phi) \;+\;\tfrac12\,\sigma^2(t)\,\nabla p_t(\phi)\Bigr]\,,$$2, and $$\partial_t p_t(\phi) =\nabla\!\cdot\!\Bigl[-f(\phi,t)\,p_t(\phi) \;+\;\tfrac12\,\sigma^2(t)\,\nabla p_t(\phi)\Bigr]\,,$$3, while diffusion-model values are $$\partial_t p_t(\phi) =\nabla\!\cdot\!\Bigl[-f(\phi,t)\,p_t(\phi) \;+\;\tfrac12\,\sigma^2(t)\,\nabla p_t(\phi)\Bigr]\,,$$4, $$\partial_t p_t(\phi) =\nabla\!\cdot\!\Bigl[-f(\phi,t)\,p_t(\phi) \;+\;\tfrac12\,\sigma^2(t)\,\nabla p_t(\phi)\Bigr]\,,$$5, and $$\partial_t p_t(\phi) =\nabla\!\cdot\!\Bigl[-f(\phi,t)\,p_t(\phi) \;+\;\tfrac12\,\sigma^2(t)\,\nabla p_t(\phi)\Bigr]\,,$$6 (Vega et al., 30 Oct 2025). The corresponding values of $$\partial_t p_t(\phi) =\nabla\!\cdot\!\Bigl[-f(\phi,t)\,p_t(\phi) \;+\;\tfrac12\,\sigma^2(t)\,\nabla p_t(\phi)\Bigr]\,,$$7 likewise agree within errors (Vega et al., 30 Oct 2025). The effective sample size for $$\partial_t p_t(\phi) =\nabla\!\cdot\!\Bigl[-f(\phi,t)\,p_t(\phi) \;+\;\tfrac12\,\sigma^2(t)\,\nabla p_t(\phi)\Bigr]\,,$$8 ensembles decreases from approximately $$\partial_t p_t(\phi) =\nabla\!\cdot\!\Bigl[-f(\phi,t)\,p_t(\phi) \;+\;\tfrac12\,\sigma^2(t)\,\nabla p_t(\phi)\Bigr]\,,$$9 to ${\rm U}(1)$00 as ${\rm U}(1)$01 grows, but resampling still yields unbiased estimates (Vega et al., 30 Oct 2025).

These results situate symmetry-preserving diffusion models in a regime beyond global discrete symmetries. The ${\rm U}(1)$02 study demonstrates that the same framework can accommodate local gauge constraints, provided that the representation and network design respect the exact transformation law.

6. Critical slowing down, interpretation, and limitations

The paper’s central interpretation is that symmetry preservation alleviates critical slowing down in lattice field theory (Vega et al., 30 Oct 2025). By enforcing exact symmetries, diffusion-generated ensembles have much larger ESS and higher Metropolis–Hastings acceptance rates than non-equivariant networks, so fewer generated configurations must be discarded (Vega et al., 30 Oct 2025). The observed drop in integrated autocorrelation times—from ${\rm U}(1)$03 for HMC to ${\rm U}(1)$04 for the tested diffusion-assisted chains—is presented as effectively eliminating critical slowing down at the volumes studied (Vega et al., 30 Oct 2025).

The proposed mechanism is partly representational. Scores, interpreted as forces, are learned more accurately because the network does not have to expend capacity discovering the symmetry (Vega et al., 30 Oct 2025). The paper further states that exact equivariance allows one to use smaller networks and fewer decoding steps for the same sample quality (Vega et al., 30 Oct 2025). This suggests a broader principle: in lattice generative modeling, architectural symmetry constraints can function simultaneously as inductive bias, variance reduction, and numerical stabilizer.

The discussion also identifies several limitations and open problems. Volume scaling remains to be studied, because ESS presently decreases with lattice size; the paper notes that advanced corrector schemes such as MALA or multiscale architectures may be needed (Vega et al., 30 Oct 2025). Extension to non-Abelian groups such as ${\rm U}(1)$05, inclusion of fermions with nonlocal determinants, and four-dimensional gauge theories are described as both theoretical and computational challenges (Vega et al., 30 Oct 2025). Further open directions include tuning of noise schedules, the choice of ${\rm U}(1)$06, integration algorithms, and network architectures such as equivariant message-passing (Vega et al., 30 Oct 2025).

A common misconception is that symmetry preservation can be replaced by ordinary data augmentation without loss. The paper directly contrasts these strategies in the gauge setting, where exact ${\rm U}(1)$07 equivariance is enforced by design rather than by data augmentation (Vega et al., 30 Oct 2025). Another potential misunderstanding is that higher-order integrators alone are responsible for improved performance; the reported trade-off indicates that RK4 yields only marginal gains over Euler at four times the number of function evaluations (Vega et al., 30 Oct 2025). Within the reported experiments, the dominant gains are associated with symmetry enforcement and force regularization rather than solver order alone.

7. Relation to diffusion-model literature

The lattice-field framework is a specialized instantiation of score-based generative modeling, especially the reverse-SDE and probability-flow formalism introduced in the general diffusion literature (Song et al., 2020). The use of denoising score matching also follows the broader score-based tradition (Song et al., 2019). Likewise, the forward noising and reverse denoising viewpoint parallels diffusion probabilistic models (Ho et al., 2020). What distinguishes the lattice-field construction is the combination of these generic ingredients with exact group equivariance and explicit force information derived from the action (Vega et al., 30 Oct 2025).

In this sense, score-based symmetry-preserving diffusion models occupy an intersection of several research programs: score-based generative modeling, equivariant neural networks, and machine-learning methods for lattice quantum field theory. The paper’s summary formulation emphasizes three elements: a reversible continuous SDE formulation of generative sampling, exact enforcement of physical symmetries by construction, and a novel force-regularized objective (Vega et al., 30 Oct 2025). Within the tested two-dimensional ${\rm U}(1)$08 and ${\rm U}(1)$09 systems, this combination yields high-quality, low-autocorrelation ensembles even near criticality (Vega et al., 30 Oct 2025).